fix: scope field extraction to target class to prevent cross-class injection

[mashraf-222]

Problem

When the LLM generates optimization code containing inner/anonymous classes with fields, those fields are incorrectly injected into the target method's enclosing class. This produces uncompilable code — e.g., type parameters like Iterator<? extends F> injected where F is not in scope.

Observed in Guava PR #12 (Iterables.mergeSorted) where inner class fields were injected into the outer Iterables class.

Root Cause

_parse_optimization_source() in replacement.py (line 72) called analyzer.find_fields(new_source) without a class_name filter, extracting ALL fields from ALL classes in the generated code — including inner classes, static nested classes, and anonymous classes.

Fix

Pass class_name=target_method.class_name to find_fields() so only fields belonging to the target method's class are extracted. The find_fields() API already supported this filter via its _walk_tree_for_fields() implementation — it just wasn't being used.

The field extraction was also moved after the target method lookup (previously it ran before we knew which class the target was in).

Validation

• Bug reproduced: test creates an inner class with badField → it was injected into outer class
• Fix confirmed: badField no longer injected, only OFFSET (same class as target) is added
• All 33 replacement tests pass (32 existing + 1 new)

Test Coverage

• tests/test_languages/test_java/test_replacement.py — TestFieldInjectionClassFiltering::test_inner_class_fields_not_injected_into_outer

Closes CF-1087

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.

[mashraf-222]

Review

Bug premise verified — real. _parse_optimization_source in codeflash/languages/java/replacement.py:72 on main calls analyzer.find_fields(new_source) without a class_name filter, so fields declared inside nested / anonymous / sibling classes in the LLM-generated optimization leak into new_fields_to_add of the outer target class, producing uncompilable code (the Guava Iterables.mergeSorted case: Iterator<? extends F> referencing an out-of-scope F).

Fix is minimal and correct. find_fields(new_source, class_name=target_class_name) uses the already-supported class filter — symmetric with find_constructors(..., class_name=...) at replacement.py:124 and find_fields(result, class_name=class_name) at replacement.py:242, both of which already filter. No new abstractions, no workaround.

Correctness edge cases: top-level siblings, static-nested, and inner classes are handled correctly by parser.py:615's existing current_class tracking. Anonymous inner classes keep fields attributed to the enclosing class (same as before — this PR doesn't change that), inherited fields from parent classes in other files are unaffected. One follow-up worth a small TODO: _walk_tree_for_fields at parser.py:615 only sets current_class for class_declaration — not enum_declaration / interface_declaration / record_declaration — so a target method inside an enum with fields would drop those fields after this change (vs. previously keeping all fields from the file). Low-probability regression; not blocking this PR.

CI blockers addressed in the last two commits:

1. prek / prek was failing on pre-existing untyped tests in test_replacement.py that the prek mypy hook surfaced because this PR touches the file. Added -> None and missing Path / JavaSupport parameter annotations on all 33 test methods + the java_support fixture.
2. unit-tests (windows-latest, 3.13) — test_help_banner.py was failing on main (Rich box-drawing characters undecodable under cp1252). Added encoding="utf-8" to both subprocess.run calls — same fix as PR fix: decode help-banner test subprocess output as UTF-8 #2120.

e2e-java java-fibonacci-nogit CF-API 500 at test generation is an infra flake unrelated to this PR. Ready for re-review.